𝗔𝗦𝗜𝟬𝟲 — 𝗠𝗲𝗺𝗼𝗿𝘆 & 𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗣𝗼𝗶𝘀𝗼𝗻𝗶𝗻𝗴

The most dangerous lie may not affect the agent today. It may wait quietly in memory until tomorrow. Attackers can plant false instructions, manipulated facts or sensitive content in long-term memory, RAG stores, conversation history or shared context. Later, the agent retrieves that poisoned information and treats it as trusted experience.

Safer memory requires:

• Provenance for every stored item
• Trust labels and integrity validation
• Separation by user, tenant, task and sensitivity
• Expiration and revalidation of old context
• Monitoring for unusual memory writes and retrievals
• A safe way to inspect, correct and delete poisoned memories

Who is allowed to write to your agent’s memory, and how is that content revalidated later?

#AgenticAI #MemoryPoisoning #RAGsecurity #AISecurity #OWASP #DataSecurity #GenAIsecurity memory poisoning, context poisoning, RAG security, vector database security, persistent prompt injection, data provenance, memory isolation, context integrity, OWASP ASI06 OWASP GenAI Security Project